Konrad Kording's ability to bridge gaps between seemingly disparate fields—utilizing data science to take new approaches to everything from brain science to prosthetics and robotics—makes him a smart pick as one of Penn's newest interdisciplinary PIK professors.
By Rob Press | Photos by Addison Geary
When Konrad Kording, PhD, rolled up to our interview on his skateboard, his fashionable glasses and red jeans were among the most obvious indicators that this was not your stereotypical Ivy League professor. Within a few minutes, he bounced a slew of ideas off of our photographer and challenged me to a pre-interview game of Guitar Hero. It set the tone quickly: His is a brain forever in need of stimulation, whether it comes in the form of an arcade game or, more commonly, through assessing data and drawing connections.
It’s the latter that brought him to the University of Pennsylvania, where his ability to bridge gaps between seemingly disparate fields—utilizing data science to take new approaches to everything from brain science to robotics—will fit right in. Kording was hired earlier this year as a Penn Integrates Knowledge (PIK) University Professor with joint appointments in the Perelman School of Medicine’s department of Neuroscience and the School of Engineering and Applied Science’s department of Bioengineering.
For Kording, there’s always another view to a given problem—and frequently, it involves stepping even further back to assess things in greater totality.
“The problem is: When we study complex systems like the brain or society, we don’t know if what we conclude is really true,” he said. “The problem is that if nobody ever tells us we are wrong about our theories, we may never fix that.”
Recently, Kording decided to test whether we might be wrong about neuroscience and the human brain. How he went about it—and the things he ascertained about the very bedrock of neuroscientific study—made waves throughout the field.
That brings us to the venue of our interview: University Family Fun Center, a traditional video arcade just off of the University of Pennsylvania campus. Small and packed to bursting with games of all different types and eras, the arcade worked as a nexus for the myriad places Kording’s neuroscientific questions can take us.
Take, for example, its out-of-order Ms. Pac-Man machine.
Wrong Like Donkey Kong
When Ms. Pac-Man was released in 1982, it was state-of-the-art. It ran on a Zilog Z80 microprocessor, which enjoyed (and in some circles, still enjoys) a heated rivalry with the MOS Technology 6502 microprocessor. In their heydays four decades ago, both were considered revolutionary. Compared to modern microprocessors, of course, they’re archaic—but for Kording and fellow study author Eric Jonas, that was part of the appeal. They were just complex enough to run something like Donkey Kong.
“The cool thing about Donkey Kong is we understand how it works,” Kording said. “We understand how a microprocessor works.”
So Kording and his team took an approach they’d typically use to study the human brain, with its 100 billion or so neurons, and applied it to the MOS Technology 6502 microprocessor, with its 3,510 transistors. The relative simplicity of the microprocessor allowed researchers to understand and manipulate the relationships between each of its many components, giving them absolute control over the entire system.
The idea is simple: If neuroscientists use these methods to investigate something as complex as the brain with the expectation that the results are useful and revelatory with regard to the brain’s function, it stands to reason that using these methods to investigate something comparatively simple that we already understand should yield results that are useful and revelatory with regard to the microprocessor’s function.
But that’s not at all how it played out.
They had hoped to apply the neuroscientific study of functional connectivity to the microprocessor. Functional connectivity is the process by which neuroscientists try and figure out which parts of the brain interact with which other parts of the brain at a given time or during a given behavior.
In the microprocessor, by controlling individual transistors and seeing how the other transistors react, they hoped to reveal the interconnected ways in which the transistors operate. What Kording and Jonas found, however, turned out to be “not very meaningful.”
When they published their paper last year, it turned a lot of heads in the field of neuroscience. If scientists are using methods that don’t fully work for the analysis of simple, linear systems, as their work suggested, then how meaningful can their inferences be when applying those methods to something as complex as the human brain?
“I think we have to have methods that are good enough to at least work on a pretty simple microprocessor,” Kording said. “We’re not asking questions about the brain. We’re asking questions about the field.”
The Ultimate Gap
In the next room over from the Ms. Pac-Man machine, among skeeball and other more conventional “games of skill,” Kording lobbed basketball after basketball toward a hoop as the bright red countdown clock ticked closer to zero. He would be the first to tell you he isn’t a very good shot, but what interests Kording is why more of us aren’t worse.
Specifically, from his original field of movement science, Kording is interested in how the brain deals with uncertainty—the fact that you can’t actually account for the exact position of your body during any given movement.
“You might believe there’s never any uncertainty,” he said. “You know where your hands are, no? But it turns out you don’t. If you’re looking at me, you aren’t looking at your hands. If I give you a task with your hand—say to touch your knee or something—you will be a little wrong. Whenever we move, we have this level of uncertainty.”
Consider this: Even the greatest basketball players in the world can’t hit free throws with 100 percent accuracy. These are 15-foot shots taken at a complete standstill, with nobody trying to defend. It’s not entirely unlike the hoop game in the arcade—yet even the very best only sink around nine out of ten shots. Given all the time in the world, even the most skilled professionals on the planet can’t do it flawlessly every single time. That’s uncertainty, in a nutshell. You can train to minimize it, but you can’t eliminate it completely because it’s inherent to our nervous and musculoskeletal systems.
You just don’t notice it, because your brain masks it so well. That’s where Kording comes in.
“A lot of my previous research has asked how the brain can be so good at this,” he said. “Your brain’s so good at it that you never even know it’s a problem.”
It’s just one example of the complexity inherent to studying the brain: Unlike in Donkey Kong, where it’s easy enough to discern hardware (the microprocessor itself) and software (the program the microprocessor is coded to run), figuring out where the hardware ends and the software begins in the brain is enormously complicated. Some parts of the field of neuroscience, according to Kording, say the brain is built as a statistical machine. Others say it’s just particularly great at learning. Piecing together where he and his lab actually fall on that continuum is one of the aims of their research.
It’s what Kording referred to as “the ultimate gap” for his team to bridge. Neuroscience as a field has a reasonable understanding of how people are inclined to behave in given situations, and it has a reasonable understanding of the hardware—the nuts and bolts of the brain itself—but how that hardware actually gives rise to those behaviors is the question he’s excited to take on through the use of data science.
Prosthetics with Precision
For Kording, understanding these intricate workings of the human brain is more than just an intellectual exercise. Piecing these things together could have a real and tremendous impact on the lives of disabled patients everywhere.
Let’s go back to the arcade for a second: specifically, to the Guitar Hero machine where Kording and I performed admirably on Pat Benatar’s “Hit Me with Your Best Shot.” Guitar Hero is a game that requires a certain level of dexterity. Your fingers have to be in the right place at the right time, hundreds of times, over the course of a given song, as you compete to play it better than your opponent.
The human brain, of course, can learn how to build the connections and develop the speed necessary to play Guitar Hero almost flawlessly. Similarly, it would be almost trivial to build a robotic hand that, when programmed to do so, could play Guitar Hero with no mistakes whatsoever. It’s drawing a direct connection between the two—creating a quick, precise, dexterous robotic prosthetic that responds perfectly to the human brain—that could change life forever for disabled individuals.
“There’s a problem we’ve worked on a lot called decoding,” Kording said. “You take the signals from the brain, and if you can figure out what the subject wants to do—if you can solve that problem—you can build prosthetic devices that you can steer with your thoughts.”
If that sounds like science fiction to you, Kording doesn’t necessarily disagree: He compared such a device to the one Luke Skywalker receives after losing his hand in “The Empire Strikes Back.” It works seamlessly, imperceptibly, like a real human hand. But while science fiction typically shows us technology we could only dream of, Kording believes we’re close to something like Skywalker’s hand being a reality. Making sense of data recorded straight out of the brain and figuring out how they relate to behavior or intent opens the door to far more responsive prosthetics, robotics, and exoskeletons.
“At some level, we have these prosthetic devices,” Kording said, pointing out examples like the BrainGate trials and a University of Pittsburgh experiment in which a monkey used a brain-controlled prosthetic device to feed itself. “They’re just really slow and inefficient and imprecise.”
The trick will be to perfect these devices. Rapid improvements in data collection should accelerate that process.
“My lab has discovered what we call Stevenson’s Law, which is that the number of simultaneously recorded neurons doubles every seven years,” Kording said, adding that our ability to collect such data is only accelerating.
Being able to use more electrodes to measure more neurons, predictably, means controlling quicker and more precise movement. According to Kording, one hundred times more electrodes could translate into movement that’s ten times faster—or, at the very least, translate into movement along many axes, giving prosthetics new degrees of freedom.
Kording’s hopes for data science research at Penn venture well beyond just what we can learn about the brain: He wants to take a crack at accelerating the entire field of medicine.
“Medicine is a very high-dimensional optimization problem,” he said. “Say I care about living healthily to an old age: At which point should I take which drugs? How should I change it as various diseases develop over time? This is a very complicated problem with lots of facets, and the world can only afford a relatively small number of randomized clinical trials every year. If we can develop ways of making progress at what helps and what doesn’t help, without requiring randomized clinical trials, we can dramatically accelerate medicine.”
When he talks about using data science to guide such seismic shifts in established fields, Kording doesn’t appear to see these goals as lofty or infeasible so much as codes he has yet to crack—levels he has yet to beat.
Audio Extra: At the Arcade with Konrad Kording
Check out these excerpts from our interview with Konrad—arcade background noises and all—along with some previously unpublished photos from our shoot!
Sign up here to subscribe to the email edition of Penn Medicine magazine.